For example, an image search system exists in which an image desired by a user (hereinafter, referred to as a desired image) is retrieved from a plurality of images stored in a large capacity recording medium (for example, a hard disk or the like).
In this image search system, the user selects an inquiry image which is most similar to the desired image from a plurality of inquiry images presented by the image search system. Further, the image search system presents an image having an image feature amount which is closest to an image feature amount of the selected inquiry image among a plurality of images to the user as the desired image which is most similar to the selected inquiry image.
However, in this image search system, in a case where a semantic gap is large, it is difficult to provide the search result which the user desires. Here, the semantic gap refers to a difference generated between an image feature amount used for searching the desired image and a concept used when a person actually recognizes the desired image.
That is, in a case where a large gap is present between the similarity of images determined by a person and the similarity of images determined by the image search system, it is difficult to provide the desired image as the search result in the image search system.
In this regard, a classifier search technique exists which searches a desired image, for example, using a classifier which performs search to match recognition that a person is capable of, in order to reduce a semantic gap. The classifier is generated in advance by performing statistic learning as teacher data of a large amount of image data obtained by adding a label indicating the similarity to the desired image by a person.
According to this classifier, it is possible to reduce the semantic gap and to obtain a search result relatively desired by a user.
However, according to the above-described classifier, in a case where a plurality of users selects the same inquiry image, the same search result is provided to all the users, which may not be a desired search result to all the plurality of users.
This is because the similarities of images determined by the plurality of different users are different from each other. That is, for example, a first user may feel that the selected inquiry image is similar to an image Q1 and an image Q2 but is not similar to an image Q3. On the other hand, a second user different from the first user may feel that the selected inquiry image is similar to the image Q1 and the image Q3 but is not similar to the image Q2.
In this regard, a fit feedback technique has been proposed which enhances the search accuracy of a desired image by generating a classifier which matches user recognition on the basis of feedback from a user, for each user (for example, refer to JP-A-10-40268, JP-A-2003-228581 and JPA-2008-276775).